QEngine 1.3.1 achieves 504 tokens/second decode throughput running Qwen3.8-Flash-Next on dual RTX 5090 GPUs, outperforming other tested engines by 1.5-1.9x in single-stream decoding. The engine matches Strata's latency on agentic prompts and features a KV-in-RAM mode that reduces overhead for resuming idle conversations. Originally designed for triple RTX 3090 configurations, the engine was successfully compiled for Blackwell architecture.